Hybrid Filter-Wrapper Approach for Feature Selection in Deceptive Consumer Review Classification

Nowadays, due to the prevailing situation of the world, people are heavily focusing on online transactions. There has been a rapid increase in online transactions and several types of data generated through such transactions during the last few years. As there is no other involvement in purchasing decisions, customers make purchasing judgments through the reviews. Therefore, not only for making purchasing decisions but also customer reviews provide valuable information regarding the products for decision-makers. By considering this as an advantage, fraudulent reviewers tend to write reviews to promote or downgrade products. Deceptive reviews can be identified via reviewer behavioural features, content-related features, or review features. But all the extracted features may not be critical for identifying deceptive. This research introduces a novel filter-wrapper hybrid approach to select optimal features to identify deceptive online customer reviews. A combination of univariate and multivariate filter methods as well as a wrapper method with the bidirectional search were used to select the features. The model was evaluated using the K-Nearest Neighbor (KNN) classifier. The proposed hybrid approach shows the highest model accuracy against the sole traditional approaches. The selected optimal features used for model building are effective as they reveal the most statistically significant features when predicting the deceptive reviews..